There are a number of problems with revoking a "was sending" message:

(1) We never make any attempt to revoke data - only kvecs contibute to
con->out_skip.  However, once the header (envelope) is written to the
socket, our peer learns data_len and sets itself to expect at least
data_len bytes to follow front or front+middle.  If ceph_msg_revoke()
is called while the messenger is sending message's data portion,
anything we send after that call is counted by the OSD towards the now
revoked message's data portion.  The effects vary, the most common one
is the eventual hang - higher layers get stuck waiting for the reply to
the message that was sent out after ceph_msg_revoke() returned and
treated by the OSD as a bunch of data bytes.  This is what Matt ran
into.

(2) Flat out zeroing con->out_kvec_bytes worth of bytes to handle kvecs
is wrong.  If ceph_msg_revoke() is called before the tag is sent out or
while the messenger is sending the header, we will get a connection
reset, either due to a bad tag (0 is not a valid tag) or a bad header
CRC, which kind of defeats the purpose of revoke.  Currently the kernel
client refuses to work with header CRCs disabled, but that will likely
change in the future, making this even worse.

(3) con->out_skip is not reset on connection reset, leading to one or
more spurious connection resets if we happen to get a real one between
con->out_skip is set in ceph_msg_revoke() and before it's cleared in
write_partial_skip().

Fixing (1) and (3) is trivial.  The idea behind fixing (2) is to never
zero the tag or the header, i.e. send out tag+header regardless of when
ceph_msg_revoke() is called.  That way the header is always correct, no
unnecessary resets are induced and revoke stands ready for disabled
CRCs.  Since ceph_msg_revoke() rips out con->out_msg, introduce a new
"message out temp" and copy the header into it before sending.

Cc: stable@vger.kernel.org # 4.0+
Reported-by: NMatt Conner <matt.conner@keepertech.com>
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Tested-by: NMatt Conner <matt.conner@keepertech.com>
Reviewed-by: NSage Weil <sage@redhat.com>

Use list_for_each_entry_safe() instead of list_for_each_safe() to
simplify the code.
Signed-off-by: NGeliang Tang <geliangtang@163.com>
[idryomov@gmail.com: nuke call to list_splice_init() as well]
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

list_next_entry has been defined in list.h, so I replace list_entry_next
with it.
Signed-off-by: NGeliang Tang <geliangtang@163.com>
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

The following bit in ceph_msg_revoke_incoming() is unsafe:

    struct ceph_connection *con = msg->con;
    if (!con)
            return;
    mutex_lock(&con->mutex);
    <more msg->con use>

There is nothing preventing con from getting destroyed right after
msg->con test.  One easy way to reproduce this is to disable message
signing only on the server side and try to map an image.  The system
will go into a

    libceph: read_partial_message ffff880073f0ab68 signature check failed
    libceph: osd0 192.168.255.155:6801 bad crc/signature
    libceph: read_partial_message ffff880073f0ab68 signature check failed
    libceph: osd0 192.168.255.155:6801 bad crc/signature

loop which has to be interrupted with Ctrl-C.  Hit Ctrl-C and you are
likely to end up with a random GP fault if the reset handler executes
"within" ceph_msg_revoke_incoming():

                     <yet another reply w/o a signature>
                                   ...
          <Ctrl-C>
    rbd_obj_request_end
      ceph_osdc_cancel_request
        __unregister_request
          ceph_osdc_put_request
            ceph_msg_revoke_incoming
                                   ...
                                osd_reset
                                  __kick_osd_requests
                                    __reset_osd
                                      remove_osd
                                        ceph_con_close
                                          reset_connection
                                            <clear con->in_msg->con>
                                            <put con ref>
                                              put_osd
                                                <free osd/con>
              <msg->con use> <-- !!!

If ceph_msg_revoke_incoming() executes "before" the reset handler,
osd/con will be leaked because ceph_msg_revoke_incoming() clears
con->in_msg but doesn't put con ref, while reset_connection() only puts
con ref if con->in_msg != NULL.

The current msg->con scheme was introduced by commits 38941f80
("libceph: have messages point to their connection") and 92ce034b
("libceph: have messages take a connection reference"), which defined
when messages get associated with a connection and when that
association goes away.  Part of the problem is that this association is
supposed to go away in much too many places; closing this race entirely
requires either a rework of the existing or an addition of a new layer
of synchronization.

In lieu of that, we can make it *much* less likely to hit by
disassociating messages only on their destruction and resend through
a different connection.  This makes the code simpler and is probably
a good thing to do regardless - this patch adds a msg_con_set() helper
which is is called from only three places: ceph_con_send() and
ceph_con_in_msg_alloc() to set msg->con and ceph_msg_release() to clear
it.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Support for message signing was merged into 3.19, along with
nocephx_require_signatures option.  But, all that option does is allow
the kernel client to talk to clusters that don't support MSG_AUTH
feature bit.  That's pretty useless, given that it's been supported
since bobtail.

Meanwhile, if one disables message signing on the server side with
"cephx sign messages = false", it becomes impossible to use the kernel
client since it expects messages to be signed if MSG_AUTH was
negotiated.  Add nocephx_sign_messages option to support this use case.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

supported_features and required_features serve no purpose at all, while
nocrc and tcp_nodelay belong to ceph_options::flags.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

We can use msg->con instead - at the point we sign an outgoing message
or check the signature on the incoming one, msg->con is always set.  We
wouldn't know how to sign a message without an associated session (i.e.
msg->con == NULL) and being able to sign a message using an explicitly
provided authorizer is of no use.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Use local variable cursor in place of &msg->cursor in
read_partial_msg_data() and write_partial_msg_data().
Signed-off-by: NShraddha Barke <shraddha.6596@gmail.com>
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

This

    struct ceph_timespec ceph_ts;
    ...
    con_out_kvec_add(con, sizeof(ceph_ts), &ceph_ts);

wraps ceph_ts into a kvec and adds it to con->out_kvec array, yet
ceph_ts becomes invalid on return from prepare_write_keepalive().  As
a result, we send out bogus keepalive2 stamps.  Fix this by encoding
into a ceph_timespec member, similar to how acks are read and written.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Reviewed-by: NYan, Zheng <zyan@redhat.com>

Only ->alloc_msg() should check data_len of the incoming message
against the preallocated ceph_msg, doing it in the messenger is not
right.  The contract is that either ->alloc_msg() returns a ceph_msg
which will fit all of the portions of the incoming message, or it
returns NULL and possibly sets skip, signaling whether NULL is due to
an -ENOMEM.  ->alloc_msg() should be the only place where we make the
skip/no-skip decision.

I stumbled upon this while looking at con/osd ref counting.  Right now,
if we get a non-extent message with a larger data portion than we are
prepared for, ->alloc_msg() returns a ceph_msg, and then, when we skip
it in the messenger, we don't put the con/osd ref acquired in
ceph_con_in_msg_alloc() (which is normally put in process_message()),
so this also fixes a memory leak.

An existing BUG_ON in ceph_msg_data_cursor_init() ensures we don't
corrupt random memory should a buggy ->alloc_msg() return an unfit
ceph_msg.

While at it, I changed the "unknown tid" dout() to a pr_warn() to make
sure all skips are seen and unified format strings.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Reviewed-by: NAlex Elder <elder@linaro.org>

Signed-off-by: NYan, Zheng <zyan@redhat.com>
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Even though it's static, con_work(), being a work func, shows up in
various stacktraces a lot.  Prefix it with ceph_.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

ceph_msgr_slab_init may fail due to a temporary ENOMEM.

Delay a bit the initialization of zero_page in ceph_msgr_init and
reorder its cleanup in _ceph_msgr_exit so it's done in reverse
order of setup.

BUG_ON() will not suffer to be postponed in case it is triggered.
Signed-off-by: NBenoît Canet <benoit.canet@nodalink.com>
Reviewed-by: NAlex Elder <elder@linaro.org>
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

addr_is_blank() should return true if family is neither AF_INET nor
AF_INET6.  This is what its counterpart entity_addr_t::is_blank_ip() is
doing and it is the right thing to do: in process_banner() we check if
our address is blank and if it is "learn" it from our peer.  As it is,
we never learn our address and always send out a blank one.  This goes
way back to ceph.git commit dd732cbfc1c9 ("use sockaddr_storage; and
some ipv6 support groundwork") from 2009.

While at at, do not open-code ipv6_addr_any() and use INADDR_ANY
constant instead of 0.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Reviewed-by: NSage Weil <sage@redhat.com>

Grab a reference on a network namespace of the 'rbd map' (in case of
rbd) or 'mount' (in case of ceph) process and use that to open sockets
instead of always using init_net and bailing if network namespace is
anything but init_net.  Be careful to not share struct ceph_client
instances between different namespaces and don't add any code in the
!CONFIG_NET_NS case.

This is based on a patch from Hong Zhiguo <zhiguohong@tencent.com>.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Reviewed-by: NSage Weil <sage@redhat.com>

From struct ceph_msg_data_cursor in include/linux/ceph/messenger.h:

bool    last_piece;     /* current is last piece */

In ceph_msg_data_next():

*last_piece = cursor->last_piece;

A call to ceph_msg_data_next() is followed by:

ret = ceph_tcp_sendpage(con->sock, page, page_offset,
                        length, last_piece);

while ceph_tcp_sendpage() is:

static int ceph_tcp_sendpage(struct socket *sock, struct page *page,
                             int offset, size_t size, bool more)

The logic is inverted: correct it.
Signed-off-by: NBenoît Canet <benoit.canet@nodalink.com>
Reviewed-by: NAlex Elder <elder@linaro.org>
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

ceph_tcp_sendpage already does the work of mapping/unmapping
the zero page if needed.
Signed-off-by: NBenoît Canet <benoit.canet@nodalink.com>
Reviewed-by: NAlex Elder <elder@linaro.org>
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

This is long overdue, and is part of cleaning up how we allocate kernel
sockets that don't reference count struct net.
Signed-off-by: N"Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

- specific con->error_msg messages (e.g. "protocol version mismatch")
  end up getting overwritten by a catch-all "socket error on read
  / write", introduced in commit 3a140a0d ("libceph: report socket
  read/write error message")
- "bad message sequence # for incoming message" loses to "bad crc" due
  to the fact that -EBADMSG is used for both

Fix it, and tidy up con->error_msg assignments and pr_errs while at it.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

This reverts commit 89baaa57.

Dirty page throttling should be sufficient for us in the general case
so there is no need to use __GFP_MEMALLOC - it would be needed only in
the swap-over-rbd case, which we currently don't support.  (It would
probably take approximately the commit that is being reverted to add
that support, but we would also need the "swap" option to distinguish
from the general case and make sure swap ceph_client-s aren't shared
with anything else.)  See ceph-devel threads [1] and [2] for the
details of why enabling pfmemalloc reserves for all cases is a bad
thing.

On top of potential system lockups related to drained emergency
reserves, this turned out to cause ceph lockups in case peers are on
the same host and communicating via loopback due to sk_filter()
dropping pfmemalloc skbs on the receiving side because the receiving
loopback socket is not tagged with SOCK_MEMALLOC.

[1] "SOCK_MEMALLOC vs loopback"
    http://www.spinics.net/lists/ceph-devel/msg22998.html
[2] "[PATCH] libceph: don't set memalloc flags in loopback case"
    http://www.spinics.net/lists/ceph-devel/msg23392.html

Conflicts:
	net/ceph/messenger.c [ context: tcp_nodelay option ]

Cc: Mike Christie <michaelc@cs.wisc.edu>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Sage Weil <sage@redhat.com>
Cc: stable@vger.kernel.org # 3.18+, needs backporting
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Acked-by: NMike Christie <michaelc@cs.wisc.edu>
Acked-by: NMel Gorman <mgorman@suse.de>

TCP_NODELAY socket option set on connection sockets,
disables Nagle’s algorithm and improves latency characteristics.
tcp_nodelay(default)/notcp_nodelay option flags provided to
enable/disable setting the socket option.
Signed-off-by: NChaitanya Huilgol <chaitanya.huilgol@sandisk.com>
[idryomov@redhat.com: NO_TCP_NODELAY -> TCP_NODELAY, minor adjustments]
Signed-off-by: NIlya Dryomov <idryomov@redhat.com>

Signed-off-by: NYan, Zheng <zyan@redhat.com>

Use kvfree() from linux/mm.h instead, which is identical.  Also fix the
ceph_buffer comment: we will allocate with kmalloc() up to 32k - the
value of PAGE_ALLOC_COSTLY_ORDER, but that really is just an
implementation detail so don't mention it at all.
Signed-off-by: NIlya Dryomov <idryomov@redhat.com>

This patch has ceph's lib code use the memalloc flags.

If the VM layer needs to write data out to free up memory to handle new
allocation requests, the block layer must be able to make forward progress.
To handle that requirement we use structs like mempools to reserve memory for
objects like bios and requests.

The problem is when we send/receive block layer requests over the network
layer, net skb allocations can fail and the system can lock up.
To solve this, the memalloc related flags were added. NBD, iSCSI
and NFS uses these flags to tell the network/vm layer that it should
use memory reserves to fullfill allcation requests for structs like
skbs.

I am running ceph in a bunch of VMs in my laptop, so this patch was
not tested very harshly.
Signed-off-by: NMike Christie <michaelc@cs.wisc.edu>
Reviewed-by: NIlya Dryomov <idryomov@redhat.com>

Commit f363e45f ("net/ceph: make ceph_msgr_wq non-reentrant")
effectively removed WQ_MEM_RECLAIM flag from ceph_msgr_wq.  This is
wrong - libceph is very much a memory reclaim path, so restore it.

Cc: stable@vger.kernel.org # needs backporting for < 3.12
Signed-off-by: NIlya Dryomov <idryomov@redhat.com>
Tested-by: NMicha Krause <micha@krausam.de>
Reviewed-by: NSage Weil <sage@redhat.com>

this allow pagelist to present data that may be sent multiple times.
Signed-off-by: NYan, Zheng <zyan@redhat.com>
Reviewed-by: NSage Weil <sage@redhat.com>

Use the more common pr_warn.

Other miscellanea:

o Coalesce formats
o Realign arguments
Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>

Determining ->last_piece based on the value of ->page_offset + length
is incorrect because length here is the length of the entire message.
->last_piece set to false even if page array data item length is <=
PAGE_SIZE, which results in invalid length passed to
ceph_tcp_{send,recv}page() and causes various asserts to fire.

    # cat pages-cursor-init.sh
    #!/bin/bash
    rbd create --size 10 --image-format 2 foo
    FOO_DEV=$(rbd map foo)
    dd if=/dev/urandom of=$FOO_DEV bs=1M &>/dev/null
    rbd snap create foo@snap
    rbd snap protect foo@snap
    rbd clone foo@snap bar
    # rbd_resize calls librbd rbd_resize(), size is in bytes
    ./rbd_resize bar $(((4 << 20) + 512))
    rbd resize --size 10 bar
    BAR_DEV=$(rbd map bar)
    # trigger a 512-byte copyup -- 512-byte page array data item
    dd if=/dev/urandom of=$BAR_DEV bs=1M count=1 seek=5

The problem exists only in ceph_msg_data_pages_cursor_init(),
ceph_msg_data_pages_advance() does the right thing.  The size_t cast is
unnecessary.

Cc: stable@vger.kernel.org # 3.10+
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@redhat.com>
Reviewed-by: NAlex Elder <elder@linaro.org>

queue_con() bumps osd ref count.  We should do the reverse when
canceling con work.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NAlex Elder <elder@linaro.org>

Add dout()s to ceph_msg_{get,put}().  Also move them to .c and turn
kref release callback into a static function.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NAlex Elder <elder@linaro.org>

It has been reported that using ZFSonLinux on rbd will result in memory
corruption. The bug report can be found here:

https://github.com/zfsonlinux/spl/issues/241
http://tracker.ceph.com/issues/7790

The reason is that ZFS will send pages with page_count 0 into rbd, which in
turns send them to tcp_sendpage. However, tcp_sendpage cannot deal with
page_count 0, as it will do get_page and put_page, and erroneously free the
page.

This type of issue has been noted before, and handled in iscsi, drbd,
etc. So, rbd should also handle this. This fix address this issue by fall back
to slower sendmsg when page_count 0 detected.

Cc: Sage Weil <sage@inktank.com>
Cc: Yehuda Sadeh <yehuda@inktank.com>
Cc: stable@vger.kernel.org
Signed-off-by: NChunwei Chen <tuxoko@gmail.com>
Reviewed-by: NIlya Dryomov <ilya.dryomov@inktank.com>

Several spots in the kernel perform a sequence like:

	skb_queue_tail(&sk->s_receive_queue, skb);
	sk->sk_data_ready(sk, skb->len);

But at the moment we place the SKB onto the socket receive queue it
can be consumed and freed up.  So this skb->len access is potentially
to freed up memory.

Furthermore, the skb->len can be modified by the consumer so it is
possible that the value isn't accurate.

And finally, no actual implementation of this callback actually uses
the length argument.  And since nobody actually cared about it's
value, lots of call sites pass arbitrary values in such as '0' and
even '1'.

So just remove the length argument from the callback, that way there
is no confusion whatsoever and all of these use-after-free cases get
fixed as a side effect.

Based upon a patch by Eric Dumazet and his suggestion to audit this
issue tree-wide.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When there is no more data, ceph_msg_data_{pages,pagelist}_advance()
should not move on to the next page.
Signed-off-by: NYan, Zheng <zheng.z.yan@intel.com>

Commit f38a5181 ("ceph: Convert to immutable biovecs") introduced
a NULL pointer dereference, which broke rbd in -rc1.  Fix it.

Cc: Kent Overstreet <kmo@daterainc.com>
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Encapsulate kmalloc vs vmalloc memory allocation and freeing logic into
two helpers, ceph_kvmalloc() and ceph_kvfree(), and switch to them.

ceph_kvmalloc() kmalloc()'s a maximum of 8 pages, anything bigger is
vmalloc()'ed with __GFP_HIGHMEM set.  This changes the existing
behaviour:

- for buffers (ceph_buffer_new()), from trying to kmalloc() everything
  and using vmalloc() just as a fallback

- for messages (ceph_msg_new()), from going to vmalloc() for anything
  bigger than a page

- for messages (ceph_msg_new()), from disallowing vmalloc() to use high
  memory
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

The check that makes sure that we have enough memory allocated to read
in the entire header of the message in question is currently busted.
It compares front_len of the incoming message with iov_len field of
ceph_msg::front structure, which is used primarily to indicate the
amount of data already read in, and not the size of the allocated
buffer.  Under certain conditions (e.g. a short read from a socket
followed by that socket's shutdown and owning ceph_connection reset)
this results in a warning similar to

[85688.975866] libceph: get_reply front 198 > preallocated 122 (4#0)

and, through another bug, leads to forever hung tasks and forced
reboots.  Fix this by comparing front_len with front_alloc_len field of
struct ceph_msg, which stores the actual size of the buffer.

Fixes: http://tracker.ceph.com/issues/5425Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Rename front_max field of struct ceph_msg to front_alloc_len to make
its purpose more clear.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Similar to userspace, don't bail with "parse_ips bad ip ..." if the
specified port is port 0, instead use port CEPH_MON_PORT (6789, the
default monitor port).
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

This updates ceph_features.h so that it has all feature bits defined in
ceph.git.  In the interim since the last update, ceph.git crossed the
"32 feature bits" point, and, the addition of the 33rd bit wasn't
handled correctly.  The work-around is squashed into this commit and
reflects ceph.git commit 053659d05e0349053ef703b414f44965f368b9f0.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

In preparation for ceph_features.h update, change all features fields
from unsigned int/u32 to u64.  (ceph.git has ~40 feature bits at this
point.)
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>